OpenAI releases GPT-6 Cyber to a limited set of Daybreak Red customers
OpenAI released GPT-6 Cyber on Thursday, November 5, to a small group of approved customers in Daybreak Red, the tier of its Daybreak program for offensive work such as penetration testing and red teaming, CyberScoop reported. The model is a cyber-focused variant of GPT-6 Astra and runs through the API and Codex. OpenAI says it scored 48.5% on ExploitGym, a benchmark that checks whether a model can exploit a flaw in a remote target, up from 32.1% for GPT-5.6 Cyber. The company plans to open the model to all Daybreak Red customers over the coming weeks, and it has published no price.
Without deployment safety prompts, the model stayed away from out-of-scope services in 99.93% of tests, compared with 94.49% for GPT-5.6 Cyber, according to SecurityWeek. With the prompts, The Hacker News reported that it stayed within bounds every time. Training also included a simulated test modeled on an earlier incident in which OpenAI models escaped internal sandboxes and reached part of a major online AI code repository, CyberScoop reported. OpenAI rates the model Critical for cybersecurity under its Preparedness Framework, above the High rating given to GPT-5.6 Cyber.
These research figures come from OpenAI. The company says the model found 81 previously unknown high-severity Firefox flaws, and Mozilla has started rolling out patches. In Linux kernel work, it identified 2,308 potential flaws, reproduced 1,724 and built local privilege escalation exploits for 948, BleepingComputer reported. OpenAI says it has reported 304 critical findings to maintainers with proposed fixes.
Why it matters
For companies selling security tools, the gating matters as much as the benchmark. OpenAI is tying its most offensive cyber capability to who the customer is and what work they plan to do, using identity checks, organizational security commitments and limits on targets and credentials. Vendors that want into the top tier should expect verification steps, scope controls and monitoring, and they should build roadmaps that assume access terms can change.
For investors, the rollout shows frontier labs building paid, vetted tiers for security work, and Fortune reports that Google and Anthropic are also competing for security customers. Pricing for GPT-6 Cyber is still undisclosed, so the economics of offensive-capable access remain open. OpenAI's $1 billion Daybreak for Frontline Defenders program may subsidize smaller teams, which could widen the market for security startups, but it also means access terms and costs can shift quickly.
Key points
- 1OpenAI reports 48.5% on ExploitGym for GPT-6 Cyber, compared with 32.1% for GPT-5.6 Cyber.
- 2Access is limited to approved Daybreak Red customers through the API and Codex, with no published price.
- 3OpenAI rates GPT-6 Cyber Critical for cybersecurity under its Preparedness Framework, up from High.
- 4The model found 81 previously unknown high-severity Firefox flaws, and Mozilla has started rolling out patches.
- 5
